OpenAI Codex becomes AI workspace
- OpenAI has pushed Codex past a coding sidekick into a desktop workspace, with a dedicated app, parallel agent threads, and shared state across app, CLI, and IDE. - The key shift is control: Codex can run local, worktree, or cloud threads, review diffs, use Git inside the app, and spawn subagents on request. - That makes governance central — admins now get workspace analytics, app-action controls, and agent permissions because these tools can act, not just answer.
OpenAI’s Codex is no longer just “the thing that writes code.” It is turning into a workspace for supervising software work across surfaces, threads, and tools. That sounds subtle, but it is a real product shift. The old mental model was pair programmer. The new one is command center — one place where you direct several agents, let some keep running, and pick the work back up later. ### What changed, exactly? The biggest change is the Codex app itself. OpenAI introduced it for macOS on February 2, 2026, then expanded availability to Windows on March 4. The app is built around separate project threads, parallel work, and long-running tasks, not a single chat box that spits out code and disappears. OpenAI’s own language is telling here — it calls the app a “command center for agents.” Does “workspace” fit better than “coding tool”? Because the app is designed to hold state and organize work, not just generate snippets. Codex can pick up session history and configuration from the CLI and IDE extension, so the same project context follows you across surfaces. OpenAI also says teams can use it across the app, editor, terminal, and cloud under the same account. That is much closer to a workspace layer than a one-off assistant. ### How do the multiple agents work? Codex can run tasks side by side in separate threads, and it supports built-in Git worktrees so agents can work on the same repository without stepping on each other. OpenAI also documents subagents — smaller worker threads you explicitly ask Codex to spawn for parallel review or implementation jobs. The important detail is that Codex orchestrates them, waits for the results, and returns a combined answer. ### Why do worktrees matter so much? Because parallel agents are only useful if their changes stay isolated. Worktrees let Codex create separate checkouts for different tasks, so one agent can test a refactor while another investigates a bug. You can
